You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Security
qs moves to 6.17.0 via the repo-wide pnpm.overrides block
(GHSA-6pq8-jhm2-2vcw,
medium: qs.stringify with arrayFormat: 'comma' throws an uncaught
TypeError on array elements carrying a non-callable own toString,
reachable as a DoS by anything serialising caller-influenced structures
into a query string). The override was chosen over per-package bumps
because qs is transitive and a per-dependency bump would leave the
other resolution paths behind. The only runtime qs in the lockfile is
now 6.17.0; the nearby 6.15.0 is @types/qs, a types package.
The two remaining Dependabot alerts — deepmerge prototype poisoning
(LOW, root and apps/mobile lockfiles) — are dismissed as tolerable
risk rather than fixed, because no patched version exists: the
latest published deepmerge (4.3.1) is inside the vulnerable range.
The dismissal records the evidence: the only dependents are dev/build
tooling (jest-config, @svgr/plugin-svgo, @expo/config) merging
repo-authored configuration, not untrusted input. To be re-checked
when upstream ships a release above 4.3.1.
Fixed
The Jira connector pulls issues from GET /rest/api/3/search/jql. Jira
Cloud removed GET /rest/api/3/search (it now answers 410 Gone), which left
the connector's alert pull failing on every poll. The request already names
the fields normalize() reads, so the issue shape it receives is unchanged.
The image stable channel guard read the registry through a buildx
template, and when the runner's buildx stopped resolving it the guard
failed open — the read returned empty with stderr discarded, "empty" was
read as "first release", and the v20.0.0 and v20.1.0 publishes each carried
image stable across the v19 → v20 application major the guard exists to
refuse, silently, under green runs. The read now goes through the registry
HTTP API (manifest → config blob → version label), exactly as the chart
half always did — which is why the chart channel held while the image
channel crossed. Three states instead of two: a 404 is the ordinary
first-release state and the channel may be taken; a tag that is present
but unreadable blocks the channel for that release, because refusing is
recoverable and a wrong crossing is not; only the deliberate promote_stable dispatch may move a blocked channel. tests/test_release_channel_tags.py pins both (the blocked-channel test
fails against the pre-fix block). image-channel.yml is the image half of
the chart's repair tool: a dispatch re-points every image's stable at a
named published release, byte-identical, with before/after evidence per
image — and was dispatched to restore the channel to v19.0.0, the recorded
owner decision.
The nightly reproducible-builds aggregator accepts the lock-only
matrix's designed skip — its own condition skips it when every
service ships an image, which has been true since the last image-less
service gained a Dockerfile, and the aggregator had read that intended
no-op as "not every service was proved", failing every scheduled run
since 2026-10-15 while the push runs stayed green behind their path
filters. The skip is accepted exactly when the discovered lock-only set
is empty; a skip while lock-only services exist is still lost coverage,
and a failed or cancelled leg on either matrix still fails.
Fixed — a fresh Windows install could not boot, and nothing could see it
install.ps1 now generates .env's secrets instead of copying the
template.Initialize-EnvFile was a plain Copy-Item .env.example .env;
that template ships ENVIRONMENT=production with SECRET_KEY, JWT_SECRET, METRICS_TOKEN and AISOC_CREDENTIAL_KEY empty, and enforce_secure_defaults refuses to start the API in production on exactly
those values — so the install ended at "These services are not running:
api" with nothing pointing back at the copy. It now runs the same scripts/ensure_env.py that make env runs, and Python 3 joins Git,
Docker, Node and pnpm as a winget-installed prerequisite because three
steps of the install are Python programs. compose-smoke.yml had already
worked around this for itself, with a comment noting that a bare copy "has
provisioned an environment no real deployment would boot".
A port already in use no longer stops the install on Windows. Test-CorePortsFree exited 3 and told the operator to edit docker-compose.yml by hand, on the two conditions most likely in this
audience's environment — an existing Postgres on 5432, an Ollama on 11434 —
while README.md promised the opposite. It now runs the same scripts/resolve_port_conflicts.py that make up runs, and every compose
call in the installer goes through one helper so a remapped stack is
addressed consistently. The closing banner reads the moved console address
from .env rather than printing a hardcoded localhost:3000.
--no-pull and --rebuild in install.sh were parsed into an array that
nothing read once the demo handoff was removed, so both were accepted and
silently ignored while --help documented them. They now set AISOC_PULL_POLICY=never and make up UP_FLAGS=--build respectively, and
the script's header describes the flow it actually runs.
Both installers linked an issue template that did not exist, under two
different names, at the moment a failing install asks for help. .github/ISSUE_TEMPLATE/installer-failure.yml now exists and both point
at it.
Changed — one way in, and no on-ramp that cannot prove itself
The devcontainer and the sandbox README no longer send people to pnpm aisoc:demo. That compose file has no ingest service, no fusion
service and Kafka disabled — its own header says not to use it to evaluate
whether AiSOC works — and both installers were rewritten to remove the
handoff, but a fresh Codespace's attach banner and the sandbox's "when
you're ready for the real stack" line were still pointing at it. Both now
say make up. A parity-gate check covers all eight on-ramp surfaces at
once, exempting blockquotes and :::caution admonitions so the warnings
that keep evaluators away from that stack can stay.
render.yaml and infra/render/ are removed. The blueprint's header
claimed a "Deploy to Render" button in README.md that has never existed,
and what it deployed was a ~$63/month stack with AISOC_DISABLE_KAFKA set
on three services and no ingest or fusion at all — a paid one-click path to
a populated console that cannot answer whether the pipeline works. Its
path filters in isolation.yml and its two .gitleaksignore exemptions go
with it. The Fly, Railway and Coolify recipes are untouched and deploy the
same lean shape; they deserve the same scrutiny.
The README's quick start now names all four on-ramps — make up, the
Windows and Unix one-liners, pip install aisoc-sandbox, and Codespaces —
where previously a Windows reader met only make up, a command their OS
does not have. Still inside the 250-line cap, paid for by reflowing.
Fixed — install documentation that described a different product
apps/docs/docs/quickstart.md advertised a machine 1.5 GB smaller than the
measured requirement (6.5 GB / ~18 GB / py3.11+ against the measured 8 GB /
20 GB / 3.9+), re-introduced the cp .env.example .env step the Makefile
eliminated, described .env.example as shipping secrets it no longer
ships, and attributed the port check to a script that does not perform it.
apps/docs/docs/installation.md said "fourteen CORE containers" (sixteen)
and told Windows users to run PowerShell as Administrator, which install.ps1's own guidance contradicts. .env.example's header said
"the four generated secrets" (fifteen).
Four surfaces still advertised Swagger at http://localhost:8000/api/docs,
which the production-class stack make up starts answers 404 on: docs/QUICK_INSTALL.md, scripts/lab.sh's banner, and the make up
transcripts in quickstart.md and deployment/walkthrough.mdx. apps/docs/docs/api/rest.md now says plainly that make up does not serve
them and names make up-dev, which does.
The sandbox README's badge said "not yet published" and its install section
said the upload was blocked on credentials, while PyPI has been serving aisoc-sandbox 0.1.0; it also claimed the production stack needs an LLM
API key, which has not been true since the bundled local model shipped.
.pr-body.md — a pull-request body draft from the v20.1.0 media close-out
(#1492) committed to the repository root by accident — is removed. Nothing
referenced it; session scratch does not belong in the tree.
Added — gates, so none of the above drifts back
tests/test_installer_parity_gate.py asserts that both installers generate
the .env secrets and that install.ps1 resolves port conflicts and
passes the port overlay to every compose call — the two holes that let the
Windows breakers ship for as long as they existed. Three existing
assertions were matching the installer's -NoLaunchhelp text rather
than its invocations, so they now read action lines with the output
cmdlets removed.
installer-smoke.yml gains a windows-latest job that dot-sources the
installer's own Initialize-EnvFile and fails if any generated secret
comes out empty. The existing Windows jobs run -Diagnose only, which
makes no changes, and a full install needs Docker Desktop — this needs
neither.
scripts/check_profile_service_counts.py reads number words as well as
digits and registers the installation page's requirements table, which is
where "fourteen" survived every digit-based check.